polars: Polars 是一个 Rust 和 Python 中的快速多线程 DataFrame 库/内存查询引擎

2023-09-20 17:15| 来源: 网络整理| 查看: 265

Documentation: Python - Rust - Node.js - R | StackOverflow: Python - Rust - Node.js - R | User Guide | Discord

Polars: Blazingly fast DataFrames in Rust, Python, Node.js, R and SQL

Polars is a DataFrame interface on top of an OLAP Query Engine implemented in Rust using Apache Arrow Columnar Format as the memory model.

To learn more, read the User Guide.

Python >>> import polars as pl >>> df = pl.DataFrame( ... { ... "A": [1, 2, 3, 4, 5], ... "fruits": ["banana", "banana", "apple", "apple", "banana"], ... "B": [5, 4, 3, 2, 1], ... "cars": ["beetle", "audi", "beetle", "beetle", "beetle"], ... } ... ) # embarrassingly parallel execution & very expressive query language >>> df.sort("fruits").select( ... "fruits", ... "cars", ... pl.lit("fruits").alias("literal_string_fruits"), ... pl.col("B").filter(pl.col("cars") == "beetle").sum(), ... pl.col("A").filter(pl.col("B") > 2).sum().over("cars").alias("sum_A_by_cars"), ... pl.col("A").sum().over("fruits").alias("sum_A_by_fruits"), ... pl.col("A").reverse().over("fruits").alias("rev_A_by_fruits"), ... pl.col("A").sort_by("B").over("fruits").alias("sort_A_by_B_by_fruits"), ... ) shape: (5, 8) ┌──────────┬──────────┬──────────────┬─────┬─────────────┬─────────────┬─────────────┬─────────────┐ │ fruits ┆ cars ┆ literal_stri ┆ B ┆ sum_A_by_ca ┆ sum_A_by_fr ┆ rev_A_by_fr ┆ sort_A_by_B │ │ --- ┆ --- ┆ ng_fruits ┆ --- ┆ rs ┆ uits ┆ uits ┆ _by_fruits │ │ str ┆ str ┆ --- ┆ i64 ┆ --- ┆ --- ┆ --- ┆ --- │ │ ┆ ┆ str ┆ ┆ i64 ┆ i64 ┆ i64 ┆ i64 │ ╞══════════╪══════════╪══════════════╪═════╪═════════════╪═════════════╪═════════════╪═════════════╡ │ "apple" ┆ "beetle" ┆ "fruits" ┆ 11 ┆ 4 ┆ 7 ┆ 4 ┆ 4 │ │ "apple" ┆ "beetle" ┆ "fruits" ┆ 11 ┆ 4 ┆ 7 ┆ 3 ┆ 3 │ │ "banana" ┆ "beetle" ┆ "fruits" ┆ 11 ┆ 4 ┆ 8 ┆ 5 ┆ 5 │ │ "banana" ┆ "audi" ┆ "fruits" ┆ 11 ┆ 2 ┆ 8 ┆ 2 ┆ 2 │ │ "banana" ┆ "beetle" ┆ "fruits" ┆ 11 ┆ 4 ┆ 8 ┆ 1 ┆ 1 │ └──────────┴──────────┴──────────────┴─────┴─────────────┴─────────────┴─────────────┴─────────────┘ SQL >>> # create a sql context >>> context = pl.SQLContext() >>> # register a table >>> table = pl.scan_ipc("file.arrow") >>> context.register("my_table", table) >>> # the query we want to run >>> query = """ ... SELECT sum(v1) as sum_v1, min(v2) as min_v2 FROM my_table ... WHERE id1 = 'id016' ... LIMIT 10 ... """ >>> ## OPTION 1 >>> # run query to materialization >>> context.query(query) shape: (1, 2) ┌────────┬────────┐ │ sum_v1 ┆ min_v2 │ │ --- ┆ --- │ │ i64 ┆ i64 │ ╞════════╪════════╡ │ 298268 ┆ 1 │ └────────┴────────┘ >>> ## OPTION 2 >>> # Don't materialize the query, but return as LazyFrame >>> # and continue in python >>> lf = context.execute(query) >>> (lf.join(other_table) ... .group_by("foo") ... .agg( ... pl.col("sum_v1").count() ... ).collect())

SQL commands can also be ran directly from your terminal using the Polars CLI:

# run an inline sql query > polars -c "SELECT sum(v1) as sum_v1, min(v2) as min_v2 FROM read_ipc('file.arrow') WHERE id1 = 'id016' LIMIT 10" # run interactively > polars Polars CLI v0.3.0 Type .help for help. > SELECT sum(v1) as sum_v1, min(v2) as min_v2 FROM read_ipc('file.arrow') WHERE id1 = 'id016' LIMIT 10;

Refer to the Polars CLI repository for more information.

Performance 🚀🚀 Blazingly fast

Polars is very fast. In fact, it is one of the best performing solutions available. See the results in DuckDB's db-benchmark.

In the TPCH benchmarks polars is orders of magnitudes faster than pandas, dask, modin and vaex on full queries (including IO).

Lightweight

Polars is also very lightweight. It comes with zero required dependencies, and this shows in the import times:

polars: 70ms numpy: 104ms pandas: 520ms Handles larger than RAM data

If you have data that does not fit into memory, polars lazy is able to process your query (or parts of your query) in a streaming fashion, this drastically reduces memory requirements so you might be able to process your 250GB dataset on your laptop. Collect with collect(streaming=True) to run the query streaming. (This might be a little slower, but it is still very fast!)

Setup Python

Install the latest polars version with:

pip install polars

We also have a conda package (conda install -c conda-forge polars), however pip is the preferred way to install Polars.

Install Polars with all optional dependencies.

pip install 'polars[all]' pip install 'polars[numpy,pandas,pyarrow]' # install a subset of all optional dependencies

You can also install the dependencies directly.

Tag Description all Install all optional dependencies (all of the following) pandas Install with Pandas for converting data to and from Pandas Dataframes/Series numpy Install with numpy for converting data to and from numpy arrays pyarrow Reading data formats using PyArrow fsspec Support for reading from remote file systems connectorx Support for reading from SQL databases xlsx2csv Support for reading from Excel files openpyxl Support for reading from Excel files with native types deltalake Support for reading from Delta Lake Tables timezone Timezone support, only needed if are on Python=1.65.

Contributing

Want to contribute? Read our contribution guideline.

Python: compile polars from source

If you want a bleeding edge release or maximal performance you should compile polars from source.

This can be done by going through the following steps in sequence:

Install the latest Rust compiler

Install maturin: pip install maturin

cd py-polars and choose one of the following:

make build-release, fastest binary, very long compile times make build-opt, fast binary with debug symbols, long compile times make build-debug-opt, medium-speed binary with debug assertions and symbols, medium compile times make build, slow binary with debug assertions and symbols, fast compile times

Append -native (e.g. make build-release-native) to enable further optimizations specific to your CPU. This produces a non-portable binary/wheel however.

Note that the Rust crate implementing the Python bindings is called py-polars to distinguish from the wrapped Rust crate polars itself. However, both the Python package and the Python module are named polars, so you can pip install polars and import polars.

Use custom Rust function in python?

Extending polars with UDFs compiled in Rust is easy. We expose pyo3 extensions for DataFrame and Series data structures. See more in https://github.com/pola-rs/pyo3-polars.

Going big...

Do you expect more than 2^32 ~4,2 billion rows? Compile polars with the bigidx feature flag.

Or for python users install pip install polars-u64-idx.

Don't use this unless you hit the row boundary as the default polars is faster and consumes less memory.

Legacy

Do you want polars to run on an old CPU (e.g. dating from before 2011)? Install pip install polars-lts-cpu. This polars project is compiled without avx target features.

Acknowledgements

Development of Polars is proudly powered by

Xomnia